Spectral alignment of stochastic gradient descent for high-dimensional classification tasks
The Annals of Applied Probability · 01.08.2025 · 35 (4) · 2767-2822
DOI: 10.1214/25-aap2187 ↗Kurz gefasst
Fachartikel in der Zeitschrift The Annals of Applied Probability, erschienen am 01.08.2025. Verfasst von Gérard Ben Arous und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 51 Literaturangaben.
Abstract (Originalsprache)
We rigorously study the relation between the training dynamics via stochastic gradient descent (SGD) and the spectra of empirical Hessian and gradient matrices. We prove that in two canonical classification tasks for multiclass high-dimensional mixtures and either 1 or 2-layer neural networks, both the SGD trajectory and emergent outlier eigenspaces of the Hessian and gradient matrices align with a common low-dimensional subspace. Moreover, in multilayer settings this alignment occurs per layer, with the final layer’s outlier eigenspace evolving over the course of training, and exhibiting rank deficiency when the SGD converges to sub-optimal classifiers. This establishes some of the rich predictions that have arisen from extensive numerical studies in the last decade about the spectra of Hessian and information matrices over the course of training in overparametrized networks.
Zitieren
APA
Arous, G. B., Gheissari, R., Huang, J., & Jagannath, A. (2025). Spectral alignment of stochastic gradient descent for high-dimensional classification tasks. The Annals of Applied Probability, 35(4), 2767-2822. https://doi.org/10.1214/25-aap2187
Vancouver
Arous GB, Gheissari R, Huang J, Jagannath A. Spectral alignment of stochastic gradient descent for high-dimensional classification tasks. The Annals of Applied Probability. 2025;35(4):2767-2822. doi:10.1214/25-aap2187
BibTeX
@article{Arous2025,
author = {Arous, Gérard Ben and Gheissari, Reza and Huang, Jiaoyang and Jagannath, Aukosh},
title = {Spectral alignment of stochastic gradient descent for high-dimensional classification tasks},
journal = {The Annals of Applied Probability},
year = {2025},
volume = {35},
number = {4},
pages = {2767--2822},
doi = {10.1214/25-aap2187}
}
EUVIMED